BixECG — Bi-mLSTM (2,522 parameters)

Official model weights for the paper "A Compact, Uncertainty-Aware mLSTM Model for Real-Time ECG Delineation on Edge Devices" (Lung et al., 2026; preprint).

BixECG performs per-sample ECG waveform delineation — classifying every sample of a single-lead ECG as No-Wave (NW), P-wave, QRS-complex, or T-wave — using a bidirectional matrix-memory LSTM (Bi-mLSTM), the mLSTM block of the xLSTM family. The selected model has only 2,522 parameters (≈10 kB), enabling real-time inference on microcontroller-class edge hardware (Arduino Portenta H7, Cortex-M7).

Model details

Architecture	Bidirectional mLSTM, 3 blocks, embedding dim 8, 1 head, conv-stem kernel 11
Parameters	2,522
Input	Single-lead ECG, 250 Hz, 300-sample R-peak-anchored window, per-window z-score
Output	Per-sample class logits over {NW, P, QRS, T}
Training data	QTDB (PhysioNet)
License	GPL-2.0-only (© Medical University of Vienna)

Performance (tolerance-aware macro-F1)

Dataset	Macro-F1	Notes
LUDB (external)	0.94	κ 0.92; QRS F1 0.98
LVAD-DB (external, pooled 2 raters)	0.91	0.94 vs engineer, 0.88 vs cardiologist
D-Heart pilot (on-device, live)	0.93	κ 0.89; 128.6 ms/beat; 22 kB SRAM

Boundary tolerances: P ±52 ms, QRS ±48 ms, T ±124 ms.

Usage

import torch
from bixecg.models.models import get_model   # pip install git+https://github.com/CellularSyntax/BixECG

cfg = {"name": "bimlstm", "params": {
    "input_dim": 1, "num_classes": 4, "embedding_dim": 8,
    "num_blocks": 3, "num_heads": 1, "conv1d_kernel_size": 11, "dropout": 0.1}}
model = get_model(cfg)
sd = torch.load("bixecg_mlstm_2522.pt", map_location="cpu")
model.load_state_dict(sd.get("model_state_dict", sd))
model.eval()

# x: (batch, 300, 1) single-lead, 250 Hz, per-window z-scored, R-peak-anchored
# logits = model(x)  -> (batch, 300, 4); labels = logits.argmax(-1)

See the code repository for preprocessing, R-peak anchoring, tolerance-aware evaluation, and uncertainty analysis.

Data availability

QTDB and LUDB are public (PhysioNet). The LVAD-DB and D-Heart pilot recordings are not distributed with this model — they are proprietary patient data under data-protection regulations, available from the corresponding author on reasonable request.

Citation

Lung D, Heute P, Marx M, Schlöglhofer T, Abart T, Moscato F, Riebandt J, Zimpfer D, Haberbusch M. A Compact, Uncertainty-Aware mLSTM Model for Real-Time ECG Delineation on Edge Devices. 2026. (Preprint; journal submission in preparation.)